Flexible protein–ligand docking with diffusion-based side-chain packing Authors: Runze Zhang, Xinyu Jiang, Duanhua Cao, Zhaokun Wang, Jie Yu, Mingan Chen, Zhehuan Fan, Xiangtai Kong, Jiacheng Xiong, Zimei Zhang, Wei Zhang, Shengkun Ni, Yitian Wang, Minda Liao, Shenghua Gao, Sulin Zhang, Mingyue Zheng Published: 2025 ABSTRACT SignificanceProtein structures are inherently dynamic, and accurate ligand-bound conformations are fundamental to structure-based drug design. However, these conformations are often challenging to obtain in practical applications, underscoring the importance to capture conformational changes within binding pockets. Most existing methods either generate static structures or depend on computationally intensive molecular dynamics simulations. Here, we couple physics-based docking with deep-learning-driven receptor flexibility modeling to characterize protein–ligand interactions. Our approach achieves high accuracy across diverse application scenarios and offers a powerful, generalizable framework for elucidating protein conformational dynamics and advancing drug discovery. Understanding protein–ligand interactions is fundamental to many areas of biological research, from elucidating cellular processes to developing new treatments for diseases (1). Proteins, the workhorses of cells, often function by interacting with small molecules or other proteins. These interactions can trigger signaling cascades, catalyze chemical reactions, or regulate gene expression (2, 3). By comprehending these interactions at a molecular level, we can gain insights into how biological systems function and how they might be modulated for therapeutic purposes (4–6). In the realm of drug discovery, this understanding takes on a more applied focus through Structure-Based Drug Design (SBDD) (7–9). Over the past few decades, SBDD has become a primary approach for developing new medications targeting specific molecular entities. The rapid advancements in structural biology and de novo protein folding algorithms have made three-dimensional protein structures more accessible than ever before (10–12). However, these static structures often fall short of meeting the complex needs of SBDD. For instance, the apo structures (proteins without bound ligands) and the holo structures with nonhomologous ligands can lead to incorrect assumptions about ligand binding modes. Moreover, not all predicted protein structures from folding algorithms are directly suitable for SBDD (13–15), partly because proteins continuously adjust their internal motions to interact with diverse molecular partners (14, 16, 17). Recently, the advent of cofolding approaches has substantially reshaped discussions surrounding protein–ligand complex prediction. These methods have demonstrated remarkable performance compared with traditional docking techniques; however, several benchmark studies have also indicated that current cofolding frameworks do not yet fully capture consistent physically grounded interaction principles (15, 18, 19). Consequently, there is a pressing need for methods that can account for the structural flexibility of protein–ligand complexes. Such approaches would not only overcome current limitations but also expand the range of potential targets for SBDD, opening new avenues for drug discovery and development. The dynamic protein–ligand binding process is commonly described by two classical mechanisms: 1) Conformational selection: A ligand selectively binds to only a few existing receptor conformations. This binding disrupts the dynamic equilibrium of receptor conformations, causing more conformations to collapse toward the selected one (20–22). 2) Induced Fit: A ligand binds to the main free conformation of the receptor, which then undergoes a conformational change to produce a more tightly bound complex conformation (23–25). Protein conformational changes during binding processes in different systems may follow different mechanisms or have both mechanisms simultaneously (26–29). Building on these classical frameworks, we employ a diffusion-based model to learn the conditional distributions of side-chain conformations under both ligand-free and ligand-bound states, thereby characterizing the conformational transitions that underlie protein–ligand recognition. Recently, deep learning has significantly improved the performance of rigid molecular docking and enabled its application in high-throughput virtual screening (30–33). However, obtaining the holo structures is often challenging, and docking into the apo structures can lead to incorrect predictions of binding modes (17, 34–36). In most cases of rigid docking failures, adjusting the side-chain direction of the apo enables successful docking of ligands. These ligands are usually positioned close to their crystal structure conformation, typically within 2.5 Å RMSD when compared to their crystal conformation (37, 38). Numerous experiments have shown that the main-chain movement from the apo to the holo is generally minimal (typically less than 1 Å). Most docking failures are due to changes in the side-chain conformation around the pocket (39–41). Therefore, a strategy that balances accuracy and complexity might involve transitioning from full-flexibility docking to accommodating only local flexibility. This is often viewed as an effective approach for flexible docking, and most of the existing flexible docking algorithms utilize rotamer library databases and heuristic searches following this strategy. However, the computational time increases exponentially with the number of flexible side-chains. Moreover, manual specification of critical side-chains is required, limiting its application in Structure Based Virtual Screening (SBVS). Examples include methods like Vinaflex (42, 43), AutoDockFR (44) (AutoDock for Flexible Receptors). In addition, IFD-MD (45) constructs a flexible docking pipeline that can slightly adjust the main-chain of the pocket region through empirical scoring functions, implicit solvent force field energies, etc., but requires more computational resources. Among deep-learning-based methods, some only consider the coarse-grained Cα of the protein, implicitly considering some protein flexibility when obtaining the ligand pose, but they fail to obtain the atomic details of the protein–ligand complex structure required for SBDD. For instance, EDM-dock (46) rebuilds complex conformation through protein–ligand distance maps. Equibind (47), E3bind (48), Uni-Mol (49) predicts ligand three-dimensional coordinates through equivariant neural networks. Additionally, DiffDock (50) simplifies the docking problem to the generation of ligand rotation, translation, and torsion angles. Recently, FlexPose (51) expanded the prediction ability of flexible docking methods on protein side-chain conformation by directly predicting the three-dimensional coordinates of the complex, while PLACER (52) unified the modeling of both protein and ligand flexibility, enabling simultaneous design of enzymes and small molecules. However, current deep-learning methods all have their own limitations. For example, they can generate unreasonable ligand conformations due to the insufficient learning of physical constraints, or spatial conflicts can occur between the ligand and the protein due to not considering the side-chain conformation. Detailed descriptions of these baseline methods, including their core algorithms and parameter settings, are provided in SI Appendix. Therefore, we urgently need a flexible method that can avoid these limitations, efficiently consider the conformational changes of protein side-chains, and directly serve both basic biological research and drug design efforts. In this study, we developed PackDock, a flexible protein modeling framework that integrates both physics-based and deep-learning-based approaches to characterize protein–ligand interactions. Our method’s core module is PackPocket, a side-chain packing model based on the diffusion model. This model samples conformations from the distributions of both “free” and “ligand bound” states of the protein, and streamlines the problem of side-chain flexibility modeling by predicting the side-chain torsional angles. To evaluate the feasibility and practicality of our method, we designed a series of experiments across multiple tasks relevant to basic biology and drug design. These experiments include side-chain recovery, flexible redocking, and cross-docking using apo, holo, and predicted folding models. These tests are designed to align with real biological scenarios and assess PackDock’s performance in various conditions. Furthermore, we validate our method through a real-world project targeting aldehyde dehydrogenase 1B1 (ALDH1B1). This application aims to demonstrate the potential of our method in both basic research and drug discovery. We also visualize the side-chain conformations generated during the process to explore PackDock’s ability to capture potential side-chain conformational changes of key amino acids. Through these comprehensive evaluations, we aim to demonstrate the practical utility of PackDock in real-world applications relevant to basic biology and drug design. We hope this study will offer a practical strategy for the scientific community to push beyond the limitations of existing structures and enhance our understanding of protein–ligand interactions in various biological contexts. Results Overview of PackDock. When knowledge about protein–ligand binding mode is unavailable, flexible docking can be more beneficial for drug design. However, conventional flexible docking algorithms rely on inefficient heuristic searches against rotamer library databases, and current deep-learning methods could lead to unreasonable complex conformations due to a lack of physical constraints (53). To address these issues, we propose a docking approach, PackDock, which combines the advantages of deep learning and conventional docking algorithms and eliminates unnecessary iterative processes. The aim is to improve the efficiency of flexible docking without compromising atomic details of the protein–ligand complex structure needed for SBDD. As illustrated in Fig. 1A, PackDock performs two-stage sampling that jointly considers the conformational distributions of proteins in both ligand-free and ligand-bound states. This design eliminates the need for extensive computational iterations typically involved in multistep search processes. During the conformational sampling phase of the protein pocket, we leverage the conditional generation advantage of diffusion models to build two separate side-chain packing models. These models allow for the simultaneous consideration of potential pocket conformations, both with and without the ligand present in the pocket. Notably, PackDock can accommodate any docking algorithms, has great adaptability, and may benefit from using other docking algorithms to enhance flexible docking performance. The specific docking protocol, including the clustering strategy for pocket conformations and definition of flexible residues, is described in SI Appendix, section S2. Overview of PackDock. (A) PackDock operates in two stages. Stage 1 uses a diffusion model PackPocket to explore empty pocket side-chain conformations. In Stage 2, the diverse conformations obtained are used as receptor for docking. The conformation of the docked ligand is utilized to model the influence of the ligand in pocket conformation during binding. This process results in dynamic receptor–ligand complex conformations. (B) PackPocket can consider the side-chain conformational space of the ligand-free state or the ligand-bound state. (C) Diffusion process of the PackPocket module. This can be done with or without a given ligand. (D) Atom Graph and Ligand Graph of the Pocket Region. The atom graph of the pocket region comprises residue nodes, each containing atomic nodes for that residue. If a ligand is present, its atoms are used as nodes. (E) Embedding Update. The information of residue nodes is updated by atomic nodes of the corresponding residue. The information from ligand atom nodes is used as input to update the corresponding residue nodes and atomic nodes. Orange edges represent the connections between atoms and their corresponding residue, gray edges represent covalent edges, and green edges represent distance edges. (F) Torsion Score. The score model updates the torsional conformation by atomic level embedding. Stage 1: The first step is to simulate the ligand-free state. Fig. 1A shows that, in the absence of a ligand, the first pocket side-chain packing model can generate N conformations of the pocket in its “free” state by using the diversity-generating ability of the diffusion model. Each pocket conformation can sample M possible protein–ligand binding modes by docking. Finally, M∗N possible complex conformations can be obtained to simulate the conformation selection process. Stage 2: The second step is to simulate the ligand-bound state. Fig. 1A shows that, in the condition of using the multiple ligand conformations obtained in the first stage, for each ligand conformation, the second pocket side-chain packing model can generate N′ conformations of the pocket in the ligand “bound” state by using the conditional generation ability of the diffusion model. Each pocket conformation can also sample M′ possible protein–ligand binding modes by docking to complete one iteration. The above process can be repeated for k iterations, finally we can obtain kNMN′M′ different complex conformations. In this work, we aim to assess PackDock’s capability to explore the side-chain conformation space of protein pockets. To reduce the influence of ligand docking algorithms on PackDock, we set M and M′=1. An evaluation of PackDock’s runtime, including the impact of sampling size and iteration number on efficiency and performance, as well as comparisons with cofolding and flexible docking methods, is presented in SI Appendix, Fig. S1. To balance the docking efficiency and accuracy, we set k=1, N and N′=6. PackPocket, the core module of PackDock, learns the side-chain conformational distributions of protein pockets in both ligand-free and ligand-bound states. Rather than viewing the side-chain prediction task as a regression task, PackPocket defines a diffusion model to treat it as a generative task. This approach provides a better basis for capturing multimodal probability distributions of protein side-chain conformations. The diffusion process uses the side-chain torsion angle score, predicted by a geometric graph neural network at each time step, to denoise a randomly initialized side-chain conformation gradually. PackPocket defines a bijective mapping between amino acid side-chains and their torsion angles (Tor1-Tor4), with specific atom-based definitions for each residue provided in SI Appendix, Table S1. The geometric graph neural network consists of three types of nodes: protein residues, protein atoms, and ligand atoms, as shown in Fig. 1D. We have defined various types of edges, such as those based on distance, covalent bonds, and relations between protein atoms and protein residues. These edges facilitate information exchange among different nodes, as shown in Fig. 1E. The exchanged information is then aggregated to the side-chain atoms, helping to denoise a randomly initialized side-chain conformation, as shown in Fig. 1F. Detailed information about our model is provided in the Methods section, and training parameters, including the optimizer settings, training duration, and hardware configuration, are detailed in SI Appendix, section S6. In summary, PackDock integrates physics-based docking algorithms with deep-learning-based protein flexibility modeling via its core module, PackPocket. This unified framework enables precise modeling of protein–ligand complex conformations while significantly improving the efficiency and reliability of flexible docking. Performance of Side-Chain Prediction in Pocket. Protein conformations are not static, especially for the ligand-binding pocket regions, where pocket breathing or allosteric pocket can lead to different side-chain conformations (16). Currently, although there are many successful protein structure prediction algorithms (1, 10, 11), they typically focus on assessing the accuracy of a single structure and are unable to generate diverse and representative protein conformations. Our model aims to capture the conformation diversity in the pocket region. In this section, we mainly discuss the ability of PackPocket to predict accurate pocket conformations while ensuring side-chain diversity. We used two independent test sets: PDBbind time-split test set (47) and the CASP13/14 test set, to evaluate the ability of PackPocket to recover pocket side-chain with or without ligands. The full list of protein targets in the CASP and PDBbind datasets used for side-chain packing evaluation is available in SI Appendix, Table S2. Following the training process, we tested side-chain conformations within 5 Å near the ligand for the PDBbind dataset, averaging 20 residues. For the CASP13/14 test set, we extracted the pocket region using the Fpocket (54) pocket-searching tool, averaging 23 residues. For comparison, we also tested several commonly used side-chain packing models/tools, including conventional methods SCWRL (55), FASPR (56) and DL-based methods AttnPacker (57), DiffPack (58). For the generative model-based packing model: DiffPack and our method, PackPocket, we performed multiple parallel runs and selected the one closest to the ground truth protein pocket conformation for statistical analysis. RMSD calculations were performed solely on side-chain atoms from the predefined pocket regions. As shown in Fig. 2 A and B and SI Appendix, Table S3, generative model-based packing model does not show a significant advantage over regression-based packing models in terms of the accuracy of single predicted conformations. This can be considered the lower bound of performance for the generative-based packaging model. However, the accuracy of generative packing models improves with the number of samples, a trend observed in both DiffPack and PackPocket. Additionally, since PackPocket uses varying numbers of pocket region crystal structures as training data and incorporates ligand condition information, the performance improvement in the pocket regions is more pronounced compared to DiffPack. With a sample size of N = 40, PackPocket achieves atomic RMSD of 0.213 Å and 0.336 Å on the CASP and PDBbind test sets, respectively. Quantitative comparisons of side-chain RMSD across all methods and sample sizes are presented in SI Appendix, Table S3. In summary, these experiments confirm PackPocket’s ability to capture accurate pocket conformations while ensuring side-chain diversity and the feasibility of PackPocket for modeling pocket flexibility. Packing and redocking experiments across Benchmarks. (A) Side-chain atom RMSD for different sampling numbers in packing experiments on the CASP dataset. The dashed line indicates the performance of the method with a sampling number of 1. (B) Side-chain atom RMSD for different sampling numbers in packing experiments on the PDBBind dataset. (C) Proportion for the flexible docking method under different side-chain RMSD thresholds in redocking experiments. (D) MAE boxplot of side-chain conformation torsion angles (x1-4) generated by PackDock and Vinaflex during docking. The box represents the interquartile range (IQR, i.e., the 25th to 75th percentiles), with the line inside indicating the median. The whiskers extend to 1.5 × IQR, and points represent outliers beyond this range. Performance of Redocking Using Holo Structures. Redocking of flexible docking algorithm refers to the process of docking with considering the flexibility of side-chain conformations in the holo. This implies that the flexible docking algorithm initially predicts a near-native side-chain conformation, and subsequently uses this conformation as a basis to predict the ligand binding pose. Therefore, in evaluating a flexible docking method, the ability to accurately predict the side-chain conformation of the holo during redocking is of primary importance. Moreover, to reduce the influence of the similarity between the proteins in the training set and the test set on the redocking performance, we used the PDBbind time-split test set (47) (no ligand and receptor overlap) for testing. We analyzed the side-chain conformation generated by PackDock. As shown in Fig. 2C, the RMSD of the side-chain conformation generated by PackDock significantly outperformed that of Vinaflex. We also performed a statistical analysis of the mean absolute error (MAE) of different torsion angles of the side-chains and found that the error rate at positions χ3-4 was higher than at χ1-2. This discrepancy primarily stems from the inherently greater flexibility of these positions. Nevertheless, PackDock outperforms conventional side-chain conformation sampling algorithm Vinaflex, as shown in Fig. 2D. Additionally, since rigid redocking uses the ground truth holo side-chain conformation, its performance potentially represents the upper limit of flexible docking. We further analyzed the ligand RMSD in the redocking experiments and found that, due to the high accuracy of PackPocket’s side-chain prediction, PackDock’s flexible docking performance was comparable to rigid docking, significantly outperforming Vinaflex using the holo, as shown in SI Appendix, Figs. S2 and S3. In summary, these redocking experiments demonstrated the ability of PackPocket to accurately predict side-chain conformations and the feasibility of PackDock’s docking process. Performance of Docking Using Apo Structures. In most real-world application situations, it is challenging to obtain the holo that binds with the ligand of interest. Often, we can only use the apo or nonhomologous ligand-induced holo for docking simulation. In these cases, treating the protein as rigid and only considering the flexibility of the ligand may have significant limitations. This section aligns with these real-world scenarios by using the apo. Previously, Zhang et al. (59) collected 32 targets’ apo from the DUD-E database (60), each with multiple holo bound with different nonhomologous ligands, and were categorized into three groups based on the C𝛼 RMSD at the ligand binding site: group 1 with 𝐶𝛼 RMSD < 0.5 Å, group 2 with 0.5 Å < C𝛼 RMSD < 1.5 Å, and group 3 with C𝛼 RMSD > 1.5 Å. We used this dataset to test the docking performance of the apo. For each method, we generated a total of 36 complex conformations and selected the one closest to the ground truth ligand binding pose for subsequent statistical analysis. Based on the previous work (61), we define the flexible region as side-chains that are within 3.5 Å from the ligand for all the conventional flexible docking algorithms, including: AutoDockflex (44), Vinaflex (42), Sminaflex (62), Gninaflex (61). Fig. 3 A and B shows that PackDock significantly outperforms existing docking methods across all targets. Table 1 further illustrates the average RMSD of ligands in different groups and the percentage of ligand RMSD < 1 Å or < 2 Å for various methods: in group 1, PackDock’s average ligand RMSD of 1.73 Å, with 70.83% of RMSD being less than 2 Å and 31.25% of RMSD being less than 1 Å. The best among other methods is DiffDock, with an average ligand RMSD of 2.14 Å, 68.75% of RMSD below 2 Å, but only 10.41% of RMSD below 1 Å. In group 2, PackDock’s average ligand RMSD of 2.52 Å, while the next best method, AutoDockflex, has only 3.31 Å. PackDock also has 47.50% of RMSD being less than 2 Å and 16.25% of RMSD being less than 1 Å, surpassing the current methods by 18.75% and 11.25%, respectively. For group 3, which has a large C𝛼 RMSD, most methods did not achieve satisfactory docking accuracy, but PackDock still had an average ligand RMSD of 2.68 Å and 41.44% of ligand RMSD being less than 2 Å. Cross-docking experiments across Benchmarks. (A) The percentage of ligand RMSD under different thresholds in all groups, when using apo. (B) The percentage of ligand RMSD < 1 Å and <2 Å for different methods in all groups, when using apo. (C) The percentage of ligand RMSD under different thresholds in DISCO, when using holo. (D) The percentage of ligand RMSD < 1 Å and <2 Å for different methods in DISCO, when using holo. (E) Percentage of system ligands with RMSD < 2 Å across similarity bins for different methods. Similarity bins were computed using the training cutoff date of June 1, 2023, (sample size denoted as n) for Boltz-2 (solid line), and September 30, 2021 (sample size denoted as m), for other cofolding methods (dashed line). Shaded areas indicate 95% CI. (F) Percentage of system ligands with RMSD < 2 Å across similarity bins for different methods. Systems with backbone RMSD > 2 Å in Boltz-2 predictions were excluded. Table 1.Docking experiments with apo structureGroup1Group2Group3All GroupsMethod%<1 Å%<2 ÅMean%<1 Å%<2 ÅMean%<1 Å%<2 ÅMean%<1 Å%<2 ÅMean Autodock 18.75 43.752.473.7520.003.332.63 30.92 2.96 5.7130.00 2.98 Smina 2.0829.162.911.2516.255.21 5.92 19.073.623.9320.003.95 Gnina 6.2539.582.751.2520.003.884.6124.343.193.9325.713.31 Vina 8.3329.162.831.2515.004.38 5.92 21.713.615.3621.073.70 Diffdock 10.41 68.75 2.14 1.25 28.75 3.351.3119.073.502.86 30.36 3.22 Autodock flex 16.6635.412.71 5.00 27.50 3.31 1.9723.023.335.3626.433.22 Smina flex 4.1643.752.663.7516.254.13 5.92 22.363.595.0024.293.58 Gnina flex 14.5847.912.43 5.00 26.253.654.6125.653.60 6.43 29.643.41 Vina flex 10.4137.502.783.7516.254.464.6122.363.735.0023.213.78 AutoDockFR 0.06.254.090.08.864.681.3111.184.250.729.674.34PackDock 31.25 70.83 1.73 16.25 47.50 2.52 7.89 41.44 2.68 14.29 48.21 2.47 Shown is the percentage of predictions with RMSD < 2 Å and RMSD < 1 Å, and the mean RMSD for all groups and the three groups separately. The last line shows our method’s performance. Best results are in bold and second best are italic in two categories. Docking experiments with apo structure Shown is the percentage of predictions with RMSD < 2 Å and RMSD < 1 Å, and the mean RMSD for all groups and the three groups separately. The last line shows our method’s performance. Best results are in bold and second best are italic in two categories. When comparing Top-1, Top-5, and best-sample metrics across methods, PackDock consistently outperforms deep-learning-based approaches like FlexPose and DiffDock, with detailed data provided in SI Appendix, Table S4. As a general framework, the performance of PackDock can be further enhanced by integrating advanced scoring functions, as demonstrated by the results of different scoring methods in SI Appendix, Fig. S4. We also analyzed the improvements made by PackDock in refining apo side-chain conformations: PackDock significantly reduces the side-chain RMSD of apo conformations compared to the original apo structures, with a breakdown of side-chain RMSD across different backbone RMSD thresholds shown in SI Appendix, Fig. S5. Additionally, PackDock maintains superior performance across ligands of varying complexities (defined by heavy atom count and rotatable bonds), with detailed trends presented in SI Appendix, Figs. S6 and S7. Visual comparisons of apo, holo, and flexible docking complexes (PackDock vs. Vinaflex) in representative systems, highlighting PackDock’s ability to resolve steric clashes, capture missing interactions, and preserve essential contacts, are provided in SI Appendix, Fig. S8. Overall, in groups 2 and 3, as the backbone RMSD increases, all methods’ performance slightly decreases. However, our method still maintains a high performance. In all groups, PackDock increased the docking success rate by 17.85% compared to all other methods. In conclusion, PackDock can effectively consider ligand-induced protein conformational changes, and accurately predict the binding pose of ligands based on the apo. Performance of Cross-Docking Using Holo Structures. Treating protein structures induced by nonhomologous ligands as rigid could result in inaccurate ligand binding poses. This section primarily discusses flexible docking tests when only nonhomologous ligands’ protein structures are available. DISCO (63) is a large-scale cross-docking test set. Given the complexity of cross-docking, previous work downsampled DISCO to obtain 7970 protein–ligand structure pairs, where no receptor was paired with its homologous ligand. These 7970 protein–ligand structure pairs were used as the cross-docking test set in this study. We generated a total of 36 complex conformations for each method and selected the one closest to the ground truth ligand binding pose for statistical analysis and defined the flexible region as side-chains that are within 3.5 Å from the ligand for all the conventional flexible docking algorithms. In Fig. 3C, the proportion of ligand poses falling below different RMSD thresholds was visualized, and PackDock’s curve is higher than the rest of the methods. Then, we separately counted the docking success rates under different RMSD threshold values, as shown in Fig. 3D and SI Appendix, Table S5. PackDock achieved 30.67% and 65.23% success rates for RMSD < 1 Å and RMSD < 2Å, respectively. The second-best method, Gnina, had 23.01% and 61.31% docking success rates, respectively. For the more stringent criterion of RMSD < 1Å, PackDock significantly improved the docking success rate, while DL-based methods, DiffDock had much lower success rates. This result is consistent with the observation of docking experiments using the apo in Fig. 3B. It should be noted that the conventional flexible docking algorithms did not show more obvious advantages than the rigid docking methods in most cases. This could be because the same protein typically does not alter much in the binding pocket when interacting with different ligands. As shown in SI Appendix, Figs. S9 and S10 and Table S6, most holo structures of the same protein have RMSD values less than 2 Å. Similar observations were also made in previous studies (61), where they found that flexible docking might increase the ligand RMSD compared to rigid docking when the binding pocket similarity was high. This is consistent with the conclusions from the redocking tests: While conventional algorithms may increase the diversity of side-chain conformations, they typically struggle to accurately identify those near the native conformations. In contrast, PackPocket has a significant advantage in the accuracy of side-chain prediction, and PackDock more comprehensively explores the side-chain conformational space by combining ligand-free and ligand-conditioned packing. We further analyzed the impact of holo backbone conformational differences on docking performance: the performance of PackDock under different backbone and pocket change thresholds, as shown in SI Appendix, Figs. S11 and S12. Since PackPocket focuses on side-chain flexibility, an increase in backbone conformational changes leads to a corresponding decrease in docking success rates. Interestingly, the performance degradation caused by pocket change RMSD in PackDock is noticeably slower than that induced by backbone change RMSD. This indirectly demonstrates the effectiveness of PackPocket in modeling side-chain flexibility. In summary, through testing on the cross-docking dataset, PackDock has demonstrated its significant advantages over conventional flexible docking methods and DL-based docking methods. It has also proven the feasibility of flexible docking when only nonhomologous ligands’ protein structures are available. Performance of Docking Using Cofolding Predicted Structures. Over the past decades, experimental methods have resolved approximately 100,000 protein structures (64), representing only a small fraction of the known protein sequence space (65). In recent years, researchers have sought to explore this vast sequence landscape using protein-folding algorithms. However, multiple studies have demonstrated that docking accuracy using AlphaFold2-predicted structures remains limited compared with crystal structures (34, 66), primarily due to the absence of ligand information and the neglect of protein dynamics. Recently, cofolding methods have incorporated ligand information during prediction, showing remarkable performance gains over conventional docking approaches and offering exciting new opportunities for structure-based drug design (SBDD). Nonetheless, recent benchmarks have also revealed notable limitations of these methods. For instance, Shen et al. (18) evaluated AlphaFold3 across 74 GPCR structures and found that, although AF3 improves backbone accuracy relative to AF2, its predicted ligand positions are often inaccurate. Masters et al. (19) demonstrated that AF3 lacks robustness to physicochemical perturbations—such as changes in molecular size, charge, polarity, and flexibility—suggesting that the model does not yet encode a consistent framework of physical interactions. Similarly, Škrinjar et al. (15) provided evidence that many apparent cofolding successes stem from training-set memorization of ligand poses, leading to significant performance drops when applied to out-of-distribution structures—raising concerns about generalization in future drug discovery applications. PackDock adopts a hybrid strategy that couples physics-based docking with deep-learning-driven receptor flexibility modeling, a combination that is particularly valuable in scenarios where memory-based generalization fails. To evaluate performance, we adopted the Runs N Poses dataset curated by Škrinjar et al. (15), restricting our evaluation to structures released after Boltz-2’s training cutoff (June 2023) (67) and stratified results by the product of pocket coverage and pose overlap (SuCOS) relative to the training set. Because the Runs N Poses protocol evaluates cofolding methods using 25 predictions per system (five seeds × five conformations), we standardized all competing flexible docking methods to also generate 25 poses, aligning with this evaluation framework. To avoid bias between crystal and cofolded backbones, we used cofolding–predicted backbones as PackDock inputs; we also evaluated Boltz-2 with pocket constraints (Boltz-2-pocket), which on average outperforms Boltz-2. Importantly, PackDock achieved substantially higher success rates in low-similarity (<60%) scenarios. In particular, when similarity <40%, PackDock outperformed Boltz-2 by approximately 20%, and Boltz-2-pocket by 10%. These low-similarity cases are among the most challenging yet most relevant to prospective drug discovery. Under high-similarity conditions, where cofolding methods approach near-perfect accuracy (>90%), PackDock performs slightly lower but remains competitive—consistent with its reliance on physics-based search strategies rather than purely data-driven memorization. Stratified analyses of docking success rates by training set similarity, ligand complexity (rotatable bonds, heavy atoms), pocket residue count, and scoring function, as well as direct comparisons between PackDock and Boltz-2 for Top-1 / Top-5 success rates, are detailed in SI Appendix, Figs. S13–S17. PackDock remains competitive under both Top-1 and Top-5 metrics. Compared with the representative traditional flexible docking algorithm Sminaflex (evaluated using accurate backbones and excluding cases with large backbone deviations, RMSD > 2 Å), PackDock consistently outperforms across all similarity ranges. While Sminaflex can rival Boltz-2 under low-similarity conditions, its performance declines sharply at high similarity, even when provided with near-native backbones—reflecting the inherent difficulty of search-based docking in navigating large conformational spaces. In contrast, PackDock demonstrates greater robustness and stability across varying structural conditions. Furthermore, we compared PackDock with recent deep-learning-based flexible docking methods, including SurfDock (33) and PLACER (52), and observed that PackDock maintained a significant performance advantage (SI Appendix, Fig. S13). When evaluated on 27 targets using AlphaFold2-predicted backbones, PackDock also outperformed all competing methods such as IFD-MD and Vinaflex, with ligand RMSD comparisons shown in SI Appendix, Fig. S18. In summary, while cofolding approaches represent an exciting new direction for SBDD, they still face fundamental limitations related to physical interpretability and generalization. Rather than viewing cofolding as a replacement for traditional docking, we propose that combining the complementary strengths of both paradigms offers a more promising path forward. PackDock embodies this synergy—providing a practical, physically grounded, and generalizable framework for structure-based drug design, and delivering significant value to the broader SBDD community. Virtual Screening for ALDH1B1 Inhibitors. To assess the potential of PackPocket in real-world drug discovery applications, we conducted a prospective experiment using the aldehyde dehydrogenase 1B1 (ALDH1B1) target. ALDH1B1 is an enzyme involved in the metabolism of aldehydes. Research has indicated that ALDH1B1 is overexpressed and plays an important role in various cancers, such as colorectal (68) and pancreatic cancers (69). Selective inhibition of ALDH1B1 has been demonstrated to significantly impede the growth of colorectal cancer spheres and patient-derived organoids. Only a few types of ALDH1B1 inhibitors have been reported, and all are currently in preclinical research. Therefore, it is essential to discover novel small molecule inhibitors targeting ALDH1B1. Initially, an enzyme kinetics assay was carried out to measure the inhibitory activity of these compounds against ALDH1B1, with IGUANA-1 (68) as a positive control. The complete virtual screening protocol, including compound clustering criteria and library details, is provided in SI Appendix, section S3. Detailed experimental protocols for ALDH1B1 expression/purification, enzymatic activity assays, protein thermal shift (PTS) assays, surface plasmon resonance (SPR) assays, and NMR assays are provided in SI Appendix, section S19. Finally, we identified five compounds, achieving a hit rate of 7.14% (Fig. 4 A and B and SI Appendix, section S20 and Table S7). These compounds demonstrated significant inhibitory activity against ALDH1B1 (Fig. 4 C and E) and markedly enhance the thermal stability of recombinant ALDH1B1 protein (Fig. 4D). PackDock identified new hits for ALDH1B1. (A) Screening protocol for ALDH1B1. (B) Chemical structures of five hit compounds identified by PackDock, along with IGUANA-1 used as a positive control. (C) Dose–response curves of five hit compounds, determined by ALDH1B1 enzyme kinetics assay. Error bars represent the mean ± SEM of three independent experiments. (D) Effects of the five hit compounds and IGUANA-1 on the thermal stability of ALDH1B1 protein, evaluated using the Protein Thermal Shift (PTS) assay. (E) A summary of IC50 values of the indicated compounds determined by enzyme kinetics assay, melting temperature differences (ΔTm) determined with PTS assay, and KD values measured by SPR assay. (F) Binding affinity measurement of 484H9 to ALDH1B1 protein using the SPR assay. The graph shows equilibrium response units plotted against 484H9 concentrations, illustrating the binding dynamics. (G) NMR analysis confirming the direct binding of 484H9 to ALDH1B1 protein. CPMG NMR spectra for 484H9 along (red) and 484H9 in the presence of ALDH1B1 protein (green). The STD spectrum for 484H9 was recorded in the presence of ALDH1B1 protein. Supplementary experimental results for ALDH1B1 include SPR binding affinity curves for hit compounds (SI Appendix, Fig. S19), inhibition curves for ALDH1B1 and control isoforms (SI Appendix, Fig. S20), inhibition rates of 70 selected compounds (SI Appendix, Fig. S21), docking scores/rankings of hits in rigid docking vs. PackDock (SI Appendix, Table S7), and a comparison of the crystal ligand pose (7mjd) with the PackDock-predicted pose of 484H9 (SI Appendix, Fig. S22). Our experimental results demonstrate that the aforementioned five compounds exhibit selectivity within the ALDH protein family, showing significant inhibitory effects exclusively on the ALDH1B1 protein among the several proteins tested. SPR assay results revealed that the binding affinities of these five compounds ranged from 0.109 to 1.67 μM (Fig. 4 E and F and SI Appendix, Fig. S19). Additionally, for 484H9, which exhibited the highest activity among the five identified compounds, we verified its binding with ALDH1B1 using NMR, and observed attenuation of 484H9 signals after incubation with recombinant ALDH1B1 (Fig. 4G). Furthermore, positive saturation transfer difference (STD) signals were detected in the STD spectrum. These results together confirm the interaction between 484H9 and ALDH1B1. Identification of Key Amino Acid Conformational Distribution. To delve deeply into the performance of PackDock in predicting side-chain conformations, we present a series of case studies. In these cases, we start with the apo, aiming to predict the holo structure after ligand binding. Fig. 5 displays the multiple side-chain conformations predicted by PackDock, and provides the corresponding apo, holo, and ground-truth ligand pose as references for comparison. The docking approach accurately identified the amino acid residues that differed substantially between the apo and the holo, and also revealed the possible conformational distribution preferences. As shown in Fig. 5A, in apo (5e8s, green)-holo (3hmm, cyan) structure, only D351 has a significant side-chain conformation difference. PackDock successfully sampled the side-chain conformation distribution of this amino acid, and identified the conformation that matches the holo. It also captured the bimodal conformation distribution of this amino acid at this site (consistent with the apo and the holo). For the rest of the amino acids, there are only slight conformational differences, and PackDock also generated compact unimodal conformation distribution for them. Similarly, as shown in Fig. 5B, F31 in apo (1pdb, green) and holo (1boz, cyan) shows significant differences in the key amino acid conformations, and PackDock accurately captured these conformational shifts. PackDock captured conformational trends. (A) In the apo (5e8s, green)-holo (3hmm, cyan) structure, D351 exhibits a significant conformational change. (B) In the apo (1pdb, green)-holo (1boz, cyan) structure, F31 exhibits a significant conformational change. Other amino acids with smaller conformational differences show a compact unimodal distribution. (C) The statistics of the side-chain torsion angles of D351, Y282, and H283 generated by PackDock for apo(5e8s)-holo(3hmm). (D) The statistics of the side-chain torsion angles of F31, F34, and Y121 generated by PackDock for apo(1pdb)-holo(1boz). Additional visualizations of side-chain conformations predicted by PackDock in representative apo-holo systems (e.g., F797 in 2ogv/3krj and W193/N220 in 4pyi/3bmw) are provided in SI Appendix, Fig. S23. We further analyzed the torsion angle distribution generated by PackDock, as shown in Fig. 5 C and D). For amino acids with substantial variations in side-chain torsional angles between apo and holo states, PackDock can sample a broad range of torsional angle distributions, as observed in D351 in apo(5e8s)-holo(3hmm) and F31 in apo(1pdb)-holo(1boz). In contrast, for amino acids with minimal changes in side-chain torsional angles between apo and holo structures, PackDock samples more concentrated torsional angle distributions, exemplified by Y282 and H283 in apo(5e8s)-holo(3hmm) and F34 and Y121 in apo(1pdb)-holo(1boz). This demonstrates that the side-chain conformations generated by PackDock are accurate and representative. This advantage can reduce the complexity of the flexible docking process, thereby improving docking success rates and enabling the precise identification of ligand-associated key amino acid residues in the apo state. We further analyzed the conformations of flexible docking complexes (SI Appendix, Fig. S8) and found that, compared to traditional rotamer library-based side-chain prediction methods, PackDock significantly reduces steric clashes between apo and holo ligand (SI Appendix, Fig. S8 A–C), captures potential interaction conformations (SI Appendix, Fig. S8 D–F), and preserves existing interaction-relevant side-chain conformations while maintaining flexibility (SI Appendix, Fig. S8 G–I). These advantages provide a feasible direction for the optimization of the lead compound stage, which is of great significance for achieving more precise SBDD. Discussion The rapid advancement of structural biology and AI-based protein folding algorithms has enabled access to the protein structures of more novel biological targets. However, these static protein structures often do not fully capture the dynamic nature of proteins in their cellular environment. This study proposes a solution, PackDock, to address the limitations of existing static structures for understanding protein–ligand interactions. Specifically, we leverage the diffusion model’s generative diversity and conditional generation capabilities to simulate protein pocket conformations in both the ligand-free and ligand-bound states. PackDock couples physics-based docking with deep-learning-driven receptor flexibility modeling, enabling mechanistic interpretation of distinct protein–ligand binding processes. We have designed various tests to assess PackDock’s ability to model different protein conformational states: 1) Side-chain packing and redocking experiments validated the ability of PackDock to predict accurate side-chain conformations of ligand binding sites. 2) Docking with apo structures, cofolding predicted structures, and cross-docking with nonhomologous ligand-induced holo structures demonstrated PackDock’s practical value in addressing challenges of protein flexibility. 3) Virtual screening of ALDH1B1 inhibitors identified molecules with nanomolar affinity with unreported scaffolds for the protein of interest, illustrating PackDock’s potential for both basic research and drug discovery applications. Overall, PackDock can process various protein structures and predict protein–ligand complex structures with high accuracy, expanding our understanding of protein–ligand interactions in biological systems. The mechanisms of biomolecular recognition are closely linked to proteins’ intrinsic dynamics. It remains a challenge to effectively account for protein flexibility in molecular interactions. PackDock employs a data-driven method to address this longstanding problem, using deep generative learning from existing static structure data to gain insights into the packing preferences of amino acid side-chains. This knowledge is crucial for understanding molecular interactions and the dynamic behavior of proteins in their cellular context. Combining recent advances in cryoelectron microscopy and protein structure prediction, we envisage that PackDock will become an essential tool for the biological research community, providing a more realistic perspective for understanding protein–ligand binding processes in both basic biology and drug design applications. Materials and Methods Model Overview. PackPocket focuses on modeling the side-chains of ligand binding sites. A protein side-chain conformation represents a specific arrangement of side-chain atoms in three-dimensional space, we can regard the protein side-chain conformation x as an element in R3×n, where n is the number of atoms. Due to the predominantly rigid nature of bond lengths, angles, and small rings within side-chain conformations, the flexibility of the side-chain primarily resides in the torsional angles of rotatable bonds. Hence, the conformational space consistent with the initial protein side-chain conformation x forms an m-dimensional submanifold Mx∈Rm, where m represents the number of rotatable bonds within the side-chain. Side-chain packing is defined as learning the probability distribution p(x) concerning the manifold Mx. Simultaneously, PackPocket defines a unique way (SI Appendix, Table S1) to determine the side-chain torsion angles of proteins, ensuring a one-to-one mapping between the torsion space and the three-dimensional conformation space. To facilitate the model learning the torsion angle distribution, we defined the rotation direction of atoms at both ends of a torsional angle by rotating only the section near the terminal end of the side-chain, which also kept the coordinates of the main-chain atoms unchanged during the sampling process. During the sampling process, PackPocket excludes Proline residues from sampling and prevents the breakage of disulfide bonds between Cysteine residues. Additionally, following pervious works (50, 70, 71), we define the score matching objective function in torsional space but handle the conformations directly in three-dimensional coordinates. Score Model. To train a score-based diffusion model (72) on protein side-chain space, we defined a bijective side-chain torsion angle definition (SI Appendix, Table S1) and constructed a scoring model, s(x,l;t), that takes as input the three-dimensional structure of the protein conformation x, the pose of the ligand l (if present), and the diffusion time coding t. The network outputs torsion angle scores invariant under SE(3) transformations for rotatable side-chain bonds near the protein pocket. Score model is constructed using the e3nn library (73). Model details are depicted in Fig. 1, where the scoring network is structured in the form of a geometric heterogeneous graph. The nodes include the protein pocket heavy atoms and ligand heavy atoms (if exist). When constructing edges, in addition to considering distance edges and covalent edges between nodes, connections are established solely between the protein heavy atom nodes and their corresponding residue nodes. To streamline the complexity of distance edges and enhance efficiency, we applied distinct cutoff values to different node types. For example, we applied a cutoff of 5 Å between atomic nodes, 30 Å between residue nodes, and a maximum of 24 neighbors per node. The residue nodes utilize the amino acid type as initial features, while receptor heavy atom nodes are assigned predefined initial features, and ligand atoms’ features are computed using RDKit and encompass: atomic number; stereochemistry; degree; formal charge; implicit valence; number of attached hydrogen atoms; number of free valence electrons; hybridization type; and whether the atom is in an aromatic ring, etc. These features are concatenated with the sine positional encoding for diffusion time (74) and edge features used radial basis function encoding for edge lengths rab. (75) These scalar features of each node and edge are transformed into a set of scalar features through a learnable MLP for initial representations as interaction layers (Each node and edge type uses a parameter-independent MLP). In the embedding update layer, we utilized initial representations of different nodes and constructed information exchange between nodes using spherical harmonics of edge vectors through tensor products. Subsequently, each node aggregated information from connected nodes via various edges. We updated node features through residual connections, as depicted in Eq. 1, achieving the objective of feature aggregation:[1]hatype←hatype⊕BN1Natype∑btype∈NatypeYr^ab⊗ψabhbtype,[2]ψab=MLPeab,ha,hb, where ha and eab represent the representation of node a and edge ab, respectively. ⊕ denotes residual connections, BN signifies equivariant batch normalization, Na refers to neighboring nodes of a, Y represents spherical harmonics functions with ℓ maximum of 2, ⊗ denotes the tensor product of irreps representations of spherical harmonics. Eq. 3 shows the different types of nodes. Distinct convolutional layers were employed for nodes of different types, as illustrated in Fig. 1. After embedding all types of atoms, we identified the rotatable side-chains and embedded their terminal and nearby atoms. To predict the m SE(3)-invariant torsion scores, we built an additional convolutional layer, inspired by previous work (50, 70), that focused on the “rotatable bonds”. Specifically, for each “rotatable bond”g = (g0, g1) and b∈V, we obtain the node embedding h and the radial basis embedding μr of the edge length, as shown in Eqs. 4, 5, and 6. Then, we constructed a convolutional filter Tg for each bond g from the tensor product of the spherical harmonics with ℓ=2 representation of the bond axis r^g, as depicted in Eq. 7. Next, we used the filter Tg to perform convolution on the neighboring nodes of the bond, obtaining representations for the “rotatable bond”, as in Eq. 8.[4]egb=MLPμrgb ∀g,b∈Eτ,[5]Eτ=g,b| g is a rotatable bond,b∈νa, Finally, a MLP is employed to predict the final torsion angle scores Torsionscore, as Eq. 9. The parameters of each MLP are not shared. Definitions of key terms used herein can be found in the SI Appendix, Tables S8 and S9. Dataset Construction. To evaluate the accuracy of PackPocket in predicting the side-chains of pocket regions under different scenarios, we used the protein dataset BC40 (76) and the protein–ligand complex dataset PDBbind (77). BC40 represents a subset derived from the PDB database, curated by filtering sequences with a 40% similarity threshold, resulting in 36,000 proteins. To focus more on modeling side-chain conformations near the binding pockets, we employed the Fpocket tool (54) to identify potential pockets within these proteins, discovering approximately 1.5 million pocket regions. Next, we obtained about 38,000 pocket regions by ranking and scoring with Fpocket. Finally, we defined the pocket graph as the amino acids within 12 Å radius from the pocket region center, and then trained the model on these graphs to predict the side-chain conformation in the empty pocket. The dataset was randomly split and tested using external test sets CASP 13 and 14. The PDBbind v2020 dataset contains 19,443 protein–ligand complexes, covering 3,890 unique receptors and 15,193 unique ligands. We excluded all complexes that the RDKit could not handle, ultimately selecting 19,119 complexes for training a side-chain conformation prediction model based on ligand conformations. Additionally, following previous work (44), we split the dataset into training, validation, and test sets according to the release time of the complex structures. For the PDBbind dataset, we defined the pocket graph as the amino acids that have at least one atom within 8 Å of the ligand position, and then trained the ligand induced pocket packing model on these graphs. Due to the abundance of hydrogen atoms in proteins and ligands, we used RDKit’s implicit hydrogen representation to construct the pocket graph. All training data underwent processing through OpenBabel. Supplementary Material Appendix 01 (PDF)